← Back to all work

Case study · By Ewan Wills, product designer · UK · Last updated

Automation · Robotics · Personal project · Doc EW-CS-03

Bolt pick & place. Camera-guided pick and place gantry

A gantry robot that finds the holes with a camera, picks up a bolt and screws it in. I built it in 2025 as a personal project, to de-risk the ideas for my CNC milling machine on something smaller and cheaper first. I designed and built all of it: the control board, the gantry, the end effector, the printed and laser-cut parts, the firmware and the vision code.

It does the job. It is also slower than you. Six bolts took it 135 seconds; a person did the same six in 72.

The pick and place gantry: belt-driven X and Y axes on steel rods over an MDF base, with the two-servo tool head in the middle
Sht 02 — Project facts

What it is, and why I built it.

ProductA gantry robot that moves bolts from one MDF part to another and screws them in, with the target holes found by a camera.
WhyA personal project, 2025. The CNC mill was going to need belt-driven axes, printed parts that hold tolerance, and my own board and firmware. This tried all of them on a smaller machine first.
MotionX and Y on timing belts over 3 mm steel rods, driven by three NEMA 17 steppers through A4988 drivers. Workspace 300 × 300 mm.
End effectorAn MG90S servo drives a rack that lowers the tool. A second MG90S, modified for continuous rotation, turns the hex bit: one part does both directions, so there is no H-bridge riding on the tool head.
VisionPython and OpenCV. Contrast boost, edge detection and a Hough circle search find the holes; a perspective warp against corner markers turns pixels into millimetres; the coordinates go to the ESP32 over serial as plain text, like 120,95.
FirmwareC++ on PlatformIO. Each move ramps the step rate up for the first three quarters of the travel and back down over the last quarter, so the head arrives without overshooting. The screwdriver runs while the rack lowers it, for as long as the descent takes.
ElectronicsMy own ESP32-WROOM-32 board with USB-C and a CH340C wired for automatic flashing, and a regulator that runs it from USB or a single LiPo cell. Five boards cost £73 from JLCPCB, assembled.
Parts cost£43.27 for one machine, from the bill of materials.
Sht 03 — The tool head

Two servos, one rack.

CAD render of the pick and place end effector: a servo-driven rack carrying a continuous-rotation servo and hex bit
End effector, CAD render. The rack lowers the tool; the second servo turns the bit.
CAD render of the gantry tool head over the workpiece with a bolt screwed into a hole
Bolt screwed in, from the final assembly render.
Sht 04 — Design for manufacture

Designing it to be printed.

Every custom part is either 3D printed or laser cut, and most of my CAD time went on making them print right first time.

i.

The base went on the laser cutter

At 300 × 300 mm it was bigger than the printer bed, so I made it flat MDF instead of splitting it into printed pieces and gluing them back together.

ii.

Small parts on purpose

The biggest printed part takes forty-five minutes on a Bambu A1, so a part that broke, or had a mistake in it, cost under an hour to replace.

iii.

Holes that slide on rods

The carriages run on 3 mm steel rod and need to slide freely without wobble. Printed flat, a hole shrinks from elephant's foot; printed upright, its top sags. I printed them upright with a small slot cut through the top of each hole, so the sag has somewhere to go and the bore keeps its size. The clearance is 0.1 mm on the diameter, from testing.

iv.

No supports

Each part is drawn to print flat in one orientation. Printed the wrong way up, a corner piece takes 48% longer and uses 63% more filament.

Sht 05 — Results

How it did.

A full run, sped up eight times. Longer video on YouTube ↗
Three bolts35 s by hand · 60 s for the machine
Six bolts72 s by hand · 135 s for the machine

So it ran at a bit under twice a person's time, which was the target I had set.

What I'd change: the belts are tensioned by hand, so slack creeps in and the bit lands slightly off the bolt head. It has no torque sensing either, so it can't tell when a bolt is seated or cross-threaded; it screws for as long as the tool takes to come down. Belt tensioners and torque sensing would deal with both. A camera on the tool head would catch the alignment that's left.

One loose end is still in the firmware. A magnet holds the bolt in the tool, and it sometimes held on after the bolt was screwed in, so the place routine lifts the tool twice.

The finished pick and place gantry on a bench, tool head lowered over a row of holes in the MDF base
The finished machine during a test run.
Grid of processing stages from the OpenCV hole-finding code: contrast-boosted, edge and threshold images of the workpiece
Stages of the hole-finding vision code on a photo of the workpiece.
Sht 06 — Questions

Quick answers.

What does the bolt pick-and-place robot do?
It finds the holes in an MDF part with a camera, picks bolts out of one part, carries them across and screws them into the other, with nobody touching it.
Why a gantry rather than a robot arm?
The bolts all sit on one flat plane, and a belt gantry reaches every point of a flat rectangle for a fraction of the price of an arm. An arm would only earn its cost on angled or curved surfaces.
Why build it?
To de-risk the ideas for my CNC milling machine on a smaller, cheaper machine first: belt-driven axes on steel rods, parts designed for printing, my own control board and firmware. It was a personal project, designed and built by me.

Need assembly automated?

Start a project ↗